Senior Harness Engineer – AIxBio

Snr Harness Engineer – AIxBio

Agentic AI Infrastructure | Onsite in San Francisco

The opportunity: build the engineering layer that turns frontier AI models into reliable scientific agents capable of carrying out complex, long-horizon drug-safety analysis.

Why apply?

This is an opportunity to help define a new engineering discipline at the intersection of agentic AI and drug development. The company already has proprietary human data, sophisticated scientific users and adoption within the pharmaceutical industry. The next challenge is making frontier models dependable enough to perform valuable scientific work over hours, not seconds. You will own the systems that make that possible.

About the company

Our client is a San Francisco-based AI x Bio company building an integrated data and AI ecosystem to improve how medicines are discovered and developed. Its long-term mission is to predict human drug outcomes more accurately than traditional preclinical methods and early clinical trials.

The company starts with high-value drug-safety problems faced by pharmaceutical scientists, builds proprietary human datasets around them, and uses those datasets to advance machine learning models and agentic infrastructure. Its technology is already being applied within pharmaceutical and biotechnology R&D workflows.

The role

The Harness Engineer will own the software, tooling and infrastructure around the company's AI agents. This is a software and systems engineering role first: you will make agent behaviour measurable, reproducible, observable and safe.

·      Own the complete agent harness, from scaffolding and tool use to runtime controls and reliability.

·      Build pipelines and storage for runtime context, agent trajectories, evaluation results and training data.

·      Create sandboxed execution environments with rapid spin-up and teardown, designed for reproducibility and deterministic evaluation.

·      Design offline evaluation suites, production-trace tests, LLM-as-judge pipelines and regression gates.

·      Work with scientists to translate expert judgement into rubrics, golden datasets and review workflows.

·      Make every model call, tool call and state transition traceable, debuggable and replayable.

·      Engineer memory, retrieval, context compaction, checkpoints and recovery for long-running agent workflows.

·      Build loop controls including retries, budget caps, stop conditions, output verification, permissions and guardrails, while supporting RL environments and rollout infrastructure.

About you

You are a strong software engineer with depth in infrastructure, platforms, data systems or developer tools. You have shipped agentic systems using LLM APIs and can talk candidly about the failure modes, debugging challenges and design decisions involved.

·      Strong Python skills and practical experience with modern infrastructure; the technology stack includes Modal, DuckDB, FastAPI, Docker, containerisation and Terraform.

·      Hands-on experience with tool-using agents, autonomous loops and production-grade evaluation or observability systems.

·      A track record of owning outcomes end to end and shipping clean, maintainable systems without waiting for a complete specification.

·      An instinct for measurement: you build the evaluation alongside the feature and treat silent infrastructure failures as a serious data-quality risk.

·      Curiosity, technical confidence and the ability to learn quickly in a field where the playbook is still being written.

Benefits

·      End-to-end ownership of a technically important platform at an ambitious, fast-moving company.

·      Work at the frontier of autonomous agents, evaluation, scientific computing and drug safety.

·      Close collaboration with expert scientists, ML researchers and sophisticated engineering colleagues.

·      A mission with tangible industry adoption and the potential to reduce reliance on animal testing.

·      The opportunity to shape core architecture and engineering standards while the discipline is still emerging.